ODE-based RK-2 decoder boosts sign language translation to 22.96 BLEU

Sign language translation has seen strong gains with Transformer architectures, but recent progress largely depends on scaling model capacity, which drives up computational costs. In a new arXiv paper, Tuğçe Kızıltepe and Hacer Yalim Keles propose a parameter-efficient alternative that improves expressiveness without inflating the model. Their work, accepted at the ACVR 2026 workshop held with ECCV 2026, focuses on iterative refinement decoders that progressively improve latent representations through multiple internal refinement steps.

Instead of using standard residual updates in these refinement steps, the authors reinterpret the process as a numerical solution to an ordinary differential equation (ODE). They replace the residual updates with higher-order integration schemes, specifically Runge-Kutta methods (RK-2 and RK-4). These methods perform multiple function evaluations per refinement step, leading to more accurate and stable representation updates. Notably, this is the first time ODE-inspired update dynamics have been applied to sign language translation, and the approach adds no extra decoder parameters.

The results are compelling. The RK-2 decoder achieves 22.96 BLEU-4 on the PHOENIX-2014-T test set and 19.34 BLEU-4 on the CSL-Daily test set, outperforming the IPSLT baseline on both benchmarks. On CSL-Daily, it does so with fewer decoder layers and fewer refinement iterations, demonstrating that better refinement dynamics can compensate for reduced model capacity. This challenges the prevailing assumption that accuracy in sign language translation requires scaling up model size.

For practitioners, this research offers a complementary alternative to conventional scaling. The efficiency gains could make high-quality sign language translation more feasible on resource-constrained devices, potentially expanding access to assistive technologies. While the paper focuses on translation, the ODE-based refinement approach could extend to other sequence-to-sequence tasks, opening new avenues for parameter-efficient Transformer designs.